Re: [dpdk-dev] [PATCH] net/virtio: fix indirect descriptors reconnection

[Maxime Coquelin]

On 9/23/21 11:21, Ding, Xuan wrote:

Hi Maxime,

-----Original Message-----
From: Ding, Xuan
Sent: Thursday, September 23, 2021 4:36 PM
To: Maxime Coquelin <maxime.coque...@redhat.com>; Xia, Chenbo
<chenbo....@intel.com>
Cc: dev@dpdk.org; Hu, Jiayu <jiayu...@intel.com>; Wang, Yinan
<yinan.w...@intel.com>; sta...@dpdk.org; Liu, Yong <yong....@intel.com>
Subject: RE: [PATCH] net/virtio: fix indirect descriptors reconnection

Hi Maxime,

-----Original Message-----
From: Maxime Coquelin <maxime.coque...@redhat.com>
Sent: Thursday, September 23, 2021 4:16 PM
To: Ding, Xuan <xuan.d...@intel.com>; Xia, Chenbo <chenbo....@intel.com>
Cc: dev@dpdk.org; Hu, Jiayu <jiayu...@intel.com>; Wang, Yinan
<yinan.w...@intel.com>; sta...@dpdk.org; Liu, Yong <yong....@intel.com>
Subject: Re: [PATCH] net/virtio: fix indirect descriptors reconnection

Hi Xuan,

On 8/19/21 07:35, Xuan Ding wrote:

Since packed indirect descriptors are added and initialized when
initializing vring, the reconnection path also needs to be considered.

Fixes: 381f39ebb78a ("net/virtio: fix packed ring indirect descricptors setup")
Cc: sta...@dpdk.org
Cc: yong....@intel.com

Signed-off-by: Xuan Ding <xuan.d...@intel.com>
---
   drivers/net/virtio/virtqueue.c | 14 ++++++++++++++
   1 file changed, 14 insertions(+)

diff --git a/drivers/net/virtio/virtqueue.c b/drivers/net/virtio/virtqueue.c
index 1f9af3c31b..47229f35c4 100644
--- a/drivers/net/virtio/virtqueue.c
+++ b/drivers/net/virtio/virtqueue.c
@@ -208,6 +208,20 @@ virtqueue_txvq_reset_packed(struct virtqueue *vq)
                        rte_pktmbuf_free(dxp->cookie);
                        dxp->cookie = NULL;
                }

It makes me think we might save quite some bytes by not allocating
indirect descriptors when feature is not negotiated, but it might have
a cost in term of performance.

Do you mean we allocate the indirect descriptors after checking the indirect
feature in virtio TX queue,
instead of allocating it at initialization?
If so, that makes sense, and hence we don't to initialize it in two places.

Sorry for replying myself, as for performance, I can have a try and test the
impact on performance.

Thanks for looking at this, but maybe that could be done later, after
having the fix merged first if you don't have the bandwidth?

Maxime

Thanks,
Xuan

+
+               struct virtio_tx_region *txr;

Don't mix declarations within code.

+               txr = txvq->virtio_net_hdr_mz->addr;
+               /* first indirect descriptor is always the tx header */
+               struct vring_packed_desc *start_dp =
+                       txr[desc_idx].tx_packed_indir;

Ditto.

Thanks, will update in next version.

Regards,
Xuan

+               vring_desc_init_indirect_packed(start_dp,
+                     RTE_DIM(txr[desc_idx].tx_packed_indir));
+               start_dp->addr = txvq->virtio_net_hdr_mem
+                       + desc_idx * sizeof(*txr)
+                       + offsetof(struct virtio_tx_region,
+                                  tx_hdr);
+               start_dp->len = vq->hw->vtnet_hdr_size;
+
        }

        vring_desc_init_packed(vq, size);

Thanks,
Maxime